Like with any tool, it’s knowing how to use it that makes a deep-learning algorithm useful, observes Albert van Bremen.
Last week, I visited a customer interested to learn more about artificial intelligence and its application in pick-and-place robots. After a quick personal introduction, I started to share some of my learnings while working for more than four years in the field of applying deep learning to high-tech systems. Somewhat proudly I explained that almost all deep-learning algorithms out there are available as open-source implementations. This means, I said, that anybody with some Python programming experience can download deep-learning models from the internet and start training. My customer promptly asked: “If everything is open and accessible to any artificial-intelligence company, how do they differentiate between themselves?”
The question took me a bit off-guard. After a short hesitation, I replied: “In the same way that a hammer and a spade are tools that are available to everybody, not everybody can make beautiful things with them. Data and algorithms are the tools of an AI engineer. Artificial-intelligence companies can set themselves apart with their experience and knowledge of applying these tools to solve engineering problems.” While my answer kept the conversation going well at that time, I needed to reflect on it later.
Having access to data and algorithms doesn’t give any guarantees that you can make deep learning work. In my company, I introduced the Friday Afternoon Experiments, something I borrowed from Phillips Research when I was working there back in 2001. Everybody in my company can spend the Friday afternoon on a topic they’re interested in and think might be relevant for the company. It encourages knowledge development, innovation and work satisfaction.
I started a Friday Afternoon Experiment myself, repeating a Deepmind project. In 2016, Deepmind created an algorithm called Alphago that was the first to defeat a professional human Go player. In a short time, the algorithm developed into the more generic Alphazero algorithm, which was trained in one day to play Go, Chess and Shogi at world champion level.
It took me over three months to get my Alphazero to work for the less complex games Connect 4 and Othello. In one day, I can now train a strong Connect 4 or Othello Alphazero player. The project took way longer than I hoped for. It made me realize that the devil of deep-learning technology really is in the details. Deep-learning algorithms learn from data. But to set up the learning process and train it successfully, you must define many so-called hyper-parameters. Small changes matter a lot, and a large part of your time can be spent on finding good hyper-parameter settings. I’m lucky to have an experienced team to discuss problems and bottlenecks.
Besides data and algorithms, compute power was a key success factor of Deepmind. To stay with the metaphor of tools, some AI companies have power tools that differentiate them from others. Companies like OpenAI, Deepmind and Meta have huge amounts of compute power available for deep-learning purposes. The AI trinity of data–algorithms–compute power defines the complexity level of the problems they can solve. If all you have is a spade, you can dig a decent hole in a day. If you have an excavator, you can dig a swimming pool within the same timeframe. Huge compute power is something not all companies have access to and this is where some AI companies can differentiate. Deepmind trained Alphago using thousands of CPUs and hundreds of GPUs. I was limited during my experiment to 64 CPU cores and 1 GPU.
If you’re searching for a solution to a standard problem, you can almost go to any artificial-intelligence startup. However, if you have a problem that hasn’t been solved before, you need more than just data, algorithms and compute power. An experienced and dedicated team makes the difference. This might seem obvious, but AI techno-babble might easily let you think otherwise. AI is teamwork!